Sequential design of multi-fidelity computer experiments: maximizing the rate of stepwise uncertainty reduction